Volume 8, No. 2, March 2017 (Special Issue) International Journal of Advanced Research in Computer Science REVIEW ARTICLE Available Online at www.ijarcs.info Cloud computing for deep learning analytics:A survey of current trends and challenges Anam Saiyeda Mansoor Ahmad Mir Department of CSE, SEST Department of CSE, SEST Jamia Hamdard, New Delhi, India Jamia Hamdard, New Delhi, India [email protected] [email protected] Abstract: Deep learning, a sub-field of machine learning is inspired by the principle of information processing in the human brain. Its applications are autonomous driving, robotics control, machine translation etc. It needs multiple training examples of the task, specialized GPU hardware, capital investment and its libraries evolve quickly, so frequent updates are needed. Cloud computing is a type of computing in which computing resources are provided on demand. Cloud is an apt choice for a platform for deep learning analytics as it provides servers, storage and networking resources. It provides scalability, processing, storage and analytics resources. Deep learning algorithms like CNN are computationally intensive for a commercial computer for large larger datasets. Cloud computing allows prevailing the processing, memory constraints of average computers, allowing computations on larger datasets. This paper discusses how cloud computing allows us to overcome the constraints of deep learning analytics on average systems and the various platforms offered by various providers. Google, NVIDIA, IBM provide us platforms for deep learning. IBM’s Rescale, Nervana Cloud , Google deep learning cloud provide full-stack hosted platform for deep learning. Keywords: Deep learning, Artificial intelligence, scalability, virtualization I. INTRODUCTION justification of the use of the cloud platform for deep learning analytics. Artificial intelligence is ubiquitous. From daily transactional tasks like online shopping to bank 2.2 Sources of information transactions to robotics every field is affected by it. In order to carry out the review, quality journals Deep learning a part of machine learning has made its were searched and the filtering was done as per the presence felt in the machine learning world. Major universally available guidelines. The papers were players like Facebook, Microsoft and Google are all taken from the databases of IEEE, Springer and ACM using it. However, for deep learning to be effective it requires huge amounts of data. Deep learning 2.3 Search Criteria architecture ensures many layers of the neural The search convention included all the papers network. “Deep” will be useful when the depth i.e. associated to analytics on the cloud. The keyword number of layers are more in number. This requires <deep learning analytics on the cloud > and <big data more storage for this large amount of data needed for analytics on the cloud> were used. The search was training. The power requirements also increase as the further refined with the keywords <Computer tasks become computationally intensive. So the Science>. Irrelevant papers related to other disciplines traditional computers may not work very effectively. were filtered out and the results are summarized Also this leads to more capital investments by the below. company. So an easier and effective way would be the Search with the keyword in the Springer database use of the services provided by the cloud for gave 286 results. The papers were ordered on the performing deep learning. analytics are discussed basis of relevance. After refining the search to include followed by the conclusion. only the papers from computer science, 178 results were obtained. These contained 152 articles, 26 II. RESEARCH METHOD Chapters, 6 conference papers. 2.1 Need for review A thorough and exhaustive review is needed to find the gaps in existing methodologies and techniques. This paper intends to summarize the existing techniques and platforms available for performing analytics using deep learning on the cloud platform. One of the purposes of the review is to study all available techniques, challenges to traditional systems, available platforms and a © 2015-19, IJARCS All Rights Reserved 68 Anam Saiyeda et al, International Journal of Advanced Research in Computer Science, 8 (2), March 2017 (Special Issue),68-72 In the ACM digital library, the keyword gave relevant papers have attempted to justify the use of 289,237 results. Refining the search andstudying cloud platform for this. abstracts gave us 5 relevant papers finally. The focus of a paper by Tsai, Chun-Wei, et al[12] was on the question that how to develop a high performance platform to efficiently analyze big data. It also attempted to design an appropriate algorithm which can find useful information from big data. Salloum et al’s[13] paper studies the analytics on the platform Apache Spark.It is a framework for big data analytics with its advanced in-memory programming model, upper-level libraries for scalable machine learning, graph analysis, streaming and structured data processing. Another paper by Middelfart, Morten[14] performed cloud based deployment of analytics and business intelligence using two different approaches The IEEE database the keyword gave 566 results. i.e an analyst specialist platform and a social platform. Refining the search and studying abstracts gave us 9 Mohammad, Atif et al [15] developed a Big Data relevant papers finally. Architecture, considering its relation with Analytics, Cloud Services and Business Intelligence. Fiore, Sandro, et al [16] proposed a cloud infrastructure for big data analytics for climate change and biodiversity. Hamdaqa, Mohammad, et al.[17] proposed a MapReduce framework for ad-hoc cloud computing. For analytics of big data new frameworks on the cloud have been proposed. Analytics on healthcare big data is a task for getting insights into a very large data set for improving the health services.Farheen [] discussed about distributed cloud computing approaches for big data. 2.4 Study selection This enormous amount of data puts a great deal of Initial filtering of papers at the first step was done stress on the write performance, scalability, on the basis of titles. Then on the basis of keywords requirement of efficient storage and meaningful the search was refined and further selection was done. processing of this data. The traditional relational The abstract ofthe remaining papers were studied. databases are insufficient for this. So a new big data After reading the complete text the final set of papers storage architecture was proposed consisting of was selected. 16 papers were included in the final set. application cluster, storage cluster to facilitate read/write/update speedup and data optimization. [19] 2.5 Data extraction A License Plate Recognition System (LPRS) was The quality of the papers was the prime proposed using deep convolutional neural network on benchmark considered for the purpose of paper a cloud platform for plate localization, character selection. For the study 16 papers were finalized. detection and segmentation. Use of bare-metal cloud servers with kernels optimized for NVIDIA GPUs, led 2.6 Research questions to an increase in the speed of training phase of the The purpose of this review is to justify the use of CNN LPDS algorithm. The paper shows the cloud platform for deep learning and present the superiority of the performance in recall, precision and various techniques available for deep learning accuracy compared to traditional LP detecting analytics on the cloud. The selection of papers was systems. [20] A paper designed deep Convolutional done according to the following research questions. Neural Networks (CNNs) to mine the deep features of The questions are as follows. cloud. [21] Work has been done to predict VM RQ1. What are the problems and challenges faced by workload in the cloud using deep learning [22], traditional systems while performing deep learning develop efficient mobile cloud system for deep analytics? learning, [23]feature extraction for 3d point cloud data RQ2. How will the use of cloud platform improve the using autoencoder[24] Deep computation models analytics done using deep learning algorithms? have been designed to offload the expensive RQ3. What are the existing techniques available for operations to the cloud. [25]Another research shows performing deep learning on the cloud? the benefits of hardware acceleration and the high performance gains on the cloud. [26]Thus a lot of III. LITERATURE REVIEW research has been done in the field of performing Research has been carried out to justify the analytics on the cloud and the various applications of performing of big data analytics on the cloud. The big data analytics via cloud[27]. Thus it is a field with datasets in deep learning can be thought of as big data a lot of research potential. as it involves large sets of images, videos, audios. The © 2015-19, IJARCS All Rights Reserved 69 Anam Saiyeda et al, International Journal of Advanced Research in Computer Science, 8 (2), March 2017 (Special Issue),68-72 IV. DEEP LEARNING neural network (ANN) with multiple hidden layers of units between the input and output layers.[3] DNNs Deep learning is an approach to AI enabling are designed as feed forward networks. Convolutional computer systems to improve with experience and deep neural networks (CNNs) are used in computer data. It represents the world as a nested hierarchy of vision and in acoustic modeling for automatic speech concepts, where every relation is defined in terms of recognition (ASR). [4] simpler concepts and more abstract representations are computed in terms of simpler ones. [1] A. Requirements It is the latest trend in machine learning being used by the major players like Facebook, Google, Microsoft Deep Learning is a computationally intensive task. etc. Google uses it in the Android Operating System's It may involve lots of operations and tasks like matrix speech recognition system, photo search for Google+ multiplications on a large scale. Thus a GPU i.e. and video recommendations in YouTube.